Addressing the Abundance Paradox Through Topology‐Guided Graph Learning for Cross‐Cohort Classification of Type 2 Diabetes

Gut microbiome‐based machine learning models for Type 2 Diabetes (T2D) classification often generalize poorly across cohorts due to population‐specific abundance variations—a practical challenge termed the abundance paradox. To capture conserved features beyond abundance, we developed DS‐GAT‐CP (Dual‐Stream Graph Attention Network enhanced by Contrastive Pre‐training), a topology‐guided framework integrating ecological network priors with supervised contrastive learning. Dual graph attention encoders, constructed from healthy and T2D‐specific co‐occurrence networks, learn disease‐relevant topological representations fused via an attention‐gating mechanism. Evaluated on six public 16S rRNA cohorts (four Chinese, two Japanese) under multiple cross‐population scenarios, DS‐GAT‐CP achieved strong predictive performance and cross‐cohort robustness, outperforming conventional machine learning and graph neural network baselines in the primary benchmark. Graph explainability analyses identified recurrent microbial interaction patterns, while topologically important taxa often lacked differential abundance, suggesting that community topology provides complementary information beyond abundance. Although further prospective validation is warranted, DS‐GAT‐CP offers a practical, interpretable strategy for incorporating ecological network information into AI models using widely available 16S rRNA data.